Purpose: Phase space (PHSP) is commonly used as a source model in Monte Carlo (MC) simulation. It is cumbersome and inconvenient to use despite the outstanding simulation accuracy. This study aims to model a biology-guided radiotherapy (BGRT) machine by a convenient and storage-saving technique using: generative adversarial networks (GAN).
Methods: The BGRT machine is a novel radiotherapy machine that irradiates target using a 64-leaf binary MLC system. Each leaf projects a 0.625 cm opening along the x direction at 85 cm SAD. Two pairs of Y jaws move simultaneously to open 1 cm or 2 cm at isocenter. A series of PHSP sources below the collimation system were generated using BEAMnrc at field sizes from 0.625 x 2 cm² to 10 x 2 cm². For each field size, a GAN model was trained from the collimated PHSP using a python tool, GAGA-PHSP, which implemented a Wasserstein GAN with three hidden layers in the generator and discriminator. The training PHSP was 0.5 GB – 9 GB in size, depending on the field size. To validate the model, dose distributions in water were simulated with PHSP and GAN sources and compared in terms of percentage depth dose (PDD) and relative planar dose.
Results: The difference of %dd(10) simulated with PHSP and GAN sources is 0.42%±0.4%. The passing rate of 1%/1mm gamma analysis of the relative dose distribution in transverse, sagittal, and coronal planes are 92.9%±4%, 96.9%±3%, and 96.2%±3.7%, respectively. The GAN-based model can generate unlimited particles and is only 40 MB in size, saving file storage by a factor up to 200. The average training time is 5 hours.
Conclusion: The GAN-based MC model can simulate the BGRT Linac accurately. The high adaptivity, storage-saving, and time-saving features make this model potentially useful in the simulation of a new Linac instead of PHSP.